Browse Records · v0.1.0
One research app · studies organized by state
Virginia study records. Open state view separately
Training experiments and Milwaukee external test · samples collected in 2014–2016
Lower MAE and RMSE are better. Methods within each comparison use identical evaluation samples. Experiments 1–3 reuse exploratory river-held-out cross-validation. Experiment 4 is an exploratory external Milwaukee test. Overall metrics stay fixed when filtering a river.
Both axes use particles/m³. Experiments 1–3 hold out each river in turn. Experiment 4 uses predictions from a separate model fit only to the original Great Lakes data, with no Milwaukee fitting or tuning.
| River / station | Date | Measured | Model prediction | Baseline | Absolute model error | Urban land (%) | Population (people/km²) | Wastewater fraction (%) | Daily-mean discharge (m³/s) | Prior 7-day precipitation (mm) |
|---|
Experiment 4 is a separate six-input ridge model: season (sine/cosine), non-event/runoff flag, urban percentage, log(1 + population density), and log(1 + prior 7-day precipitation). It excludes discharge and wastewater contribution rather than filling missing inputs. It was fit only to the 105 original training samples; its alpha 10 and preprocessing were fixed before this run. Its baseline is the original training-data median, not the Milwaukee median. This does not validate the full experiment 3 model.
Milwaukee evaluation uses 19 surface-water samples from 2016. An ambiguous jar label was excluded. Six stations belong to three related Great Lakes river systems, so this is not broad geographic validation. The outcomes had been inspected during eligibility review: this test is exploratory, not prospectively blinded. Do not tune against these same outcomes. External R² is strongly negative and the model performs worse than its baseline.
External sample source: USGS Milwaukee 2016 release. External land-cover/population source: Lenaker et al. (2019), supporting Table S1, with its CC BY-NC 4.0 terms. Prior model descriptions below apply to experiments 1–3.
All fixed ridge regressions use season and low-flow/runoff-event condition. Experiment 2 also uses urban land percentage, log(1 + population density), and annual-mean wastewater fraction. Experiment 3 additionally uses log(1 + daily-mean discharge) and log(1 + preceding 7-day precipitation). The 7-day window was specified before running; other windows were not searched. The model standardizes inputs within training folds, fits log(1 + concentration), and uses alpha 10. Each river was held out once (29 rivers in experiments 1 and 2; 27 in experiment 3). There was no hyperparameter tuning. The baseline used the median of each training fold. Concentrations were calculated from counts divided by sampled water volume.
Five blank samples were excluded. Two environmental samples with a different lower size cutoff were excluded, leaving 105. Experiment 3 excludes another six samples without exact-station discharge records, leaving 99 samples. All comparator methods were re-fit on the same 99 samples. Six retained discharge values are flagged ESTIMATED by USGS. The target uses source size categories 0.333–1 mm and 1–4.75 mm. Larger debris was excluded. No additional blank correction was applied.
These results do not establish accuracy for Maryland, Virginia, other methods, or future conditions. No calibrated prediction intervals are available. Watershed land cover is from 2011, population from 2010, and wastewater contributions are annual means, not sampling-day discharge. The original experiments represent 29 independent watersheds; experiment 3 uses 27. Both follow-up experiments reuse historical data and require independent validation. Daily-mean discharge includes the full sampling day and can include measurements after sample collection; this is retrospective analysis, not a forecast. Daymet gives local 1-km total precipitation including rain and snow water equivalent, not watershed-average rainfall. Precipitation sums exclude the sampling day. No missing discharge was filled with zero or another gauge.
Sample data: Baldwin, Corsi and Mason (2016), USGS data release. Associated study. Watershed inputs: Baldwin, Corsi and Mason, SI Table 1, CC BY-NC 4.0. Extracted, joined and modeled for noncommercial research; not endorsed by the authors. Hydrology sources: USGS daily mean discharge and Thornton et al. (2022), Daymet V4 R1. This panel contains our derived experimental results and does not reproduce source figures.
Separate model · 105 samples · 29 rivers · exploratory leave-one-river-out evaluation
0.333mm_1mm, while the paper describes a 355 µm analytical sieve. The numerical lower-cutoff discrepancy remains unresolved. These results do not establish readiness for prediction use.| Method | MAE (particles/m³) | RMSE (particles/m³) | R² |
|---|---|---|---|
| Small-fraction ridge | 2.369 | 4.265 | 0.056 |
| Training-fold median | 2.536 | 4.778 | -0.185 |
Lower MAE and RMSE are better. The modest MAE improvement is 6.59%; R² is 0.056. The model has lower MAE at 15 of 29 held-out rivers. These scores concern a different particle-size target from experiments 1–4 and are not directly comparable with those scores. Overall metrics stay fixed when filtering a river.
| River / station | Date | Measured small fraction | Held-out model prediction | Training-fold median baseline | Absolute model error |
|---|
Only the source small-fraction count divided by sampled water volume is used as the target. The existing 105-event Great Lakes cohort excludes five field blanks and two environmental samples with a 0.125 mm lower-cutoff category. No additional blank subtraction was applied. Each source count total was checked against its five morphology counts. All 105 sample joins reconcile with the previously assembled cohort.
Fixed ridge alpha 10 uses season sine/cosine (day of year / 365.25), runoff-event indicator, urban land percentage, log(1 + population density) and log(1 + prior-seven-day precipitation). Standardization and fitting use each training fold only. The target is log(1 + concentration); inverse predictions are clipped at zero, without bias correction. All samples from one river are held out together in each fold. The baseline is the median of that training fold. No hyperparameter search was performed. These reuse previously examined Great Lakes records and provide exploratory internal evaluation.
No Northeast or Milwaukee records were used for this experiment. No Northeast predictions were generated. Watershed approximations, population-method equivalence and laboratory-method differences still need review. A final fit on all 105 records is available in the separate model package; the displayed results are held-out predictions, not final-fit training predictions.
Great Lakes source counts and metadata · Published methods. The raw source label is retained in CSV exports. Previously derived watershed inputs use supporting Table 1 under CC BY-NC 4.0; rainfall uses Daymet V4 R1. These are research results, not water-safety assessments.
When comparing concentrations, check the polymer type, particle size range, and surface condition (such as weathering) when the study reports them. Two samples with the same particle count per liter can contain different kinds of particles. Different minimum particle sizes can also affect which particles each study counts.
In laboratory tests, Qian and colleagues found that tested microplastics generated reactive oxygen species at their surfaces in water. More water-repellent surfaces showed greater activity under the study conditions. River particle counts alone cannot tell us how reactive those particles are.
Check these details in each study before comparing concentrations:
PolyGone laboratory methods · Interlaboratory study on microplastic measurements
68 size-fraction records · 34 sampling events · 17 stations. 61 records match totals and particle categories; 7 need verification (including one category-only disagreement). Count consistency does not establish model suitability. These records have not been used for model training.
Concentrations are derived from published summary counts / filtered volume (m³). No blank subtraction applied. Flagged records retain unresolved source values.
| Station | Date | Conditions | Size (µm) | Summary count | Volume (m³) | Particles/m³ | Process-blank fibers | Count check | Review note |
|---|
Same station and size fraction only. Dates are shown because these are separate sampling events. Where multiple events exist, all combinations are listed; rows are not independent replicates. This descriptive comparison does not establish a storm effect. State, station and status filters also apply here. Differences are withheld for flagged records.
| Station | Size (µm) | Baseflow date | Baseflow particles/m³ | Stormflow date | Stormflow particles/m³ | Difference particles/m³ | Comparison status |
|---|
Surface-water sampling with 333 µm mesh nets. Reported fractions: 355–1000 and 1000–5600 µm (the upper fraction extends beyond 5 mm). Visual stereomicroscopy up to 40×; spectroscopic confirmation is not documented in the supplied metadata. Fiber clumps are excluded from summary totals. Split corrections account for subsampling, not blank subtraction. Process blanks monitor airborne fibers during counting.
Four Hudson River records have possible date exchanges; two Lick Run records have a station-ID discrepancy. Original values are preserved. No coordinates have been inferred.
31 small-fraction events passed count checks · all 31 have preceding-seven-day precipitation · 21 have three inputs available, including 8 events with approximate watershed inputs · 10 have missing or withheld watershed inputs. Availability does not establish model readiness.
Review pending — no Northeast predictions
355–1000 µm source fraction. Three source-discrepancy events are excluded from this input table. Six events have separate candidate watershed values; these do not count toward the 21 available-input events. Peconic and Bound Brook remain on hold. Three precipitation totals exceed the original training maximum of 75.77 mm.
This table has its own filter, independent of the record filters above. Candidate values remain separate from available inputs. Blank counts are shown for context; no blank subtraction was applied.
| Station / ID | Date | Developed land (%) · 2011 | Population (people/km²) · 2010 | Prior 7-day precipitation (mm) | Input status | Watershed review | Predictor range checks | Fiber clumps / process-blank fibers | Separate candidate developed (%) / population |
|---|
The 34 small-fraction events include 31 exact-ID/date/fraction count matches and three source discrepancies held out of this input table. Counts matching is not polymer confirmation. Fiber clumps are excluded from source summary totals and retained as flags; clump counts do not represent individual fibers. The full study table above retains all 68 size-fraction records and source values.
Published GAGES-II attributes cover 13 events; approximate StreamCat attributes cover eight. GAGES-II source IDs are matched exactly. Developed land means NLCD 2011 classes 21–24, not impervious cover. Population refers to Census 2010. Basin boundaries remain qualified. GAGES-II source tables · StreamCat definitions.
Separate Assunpink, Codorus and Rock Creek candidate values cover six events. Their whole-watershed area differences from published USGS station areas are +0.67%, +0.23% and +3.95%. Station points fall within the whole watersheds, but exact-outlet split requests returned service errors. Candidate values remain unaccepted. Peconic lacks an independent area check; Bound Brook retains its area discrepancy. Their values remain withheld.
A six-station source comparison found population differences from −4.84% to +16.22% and developed-land differences up to 4.68 percentage points. Differences confound watershed boundaries and aggregation methods; StreamCat and GAGES-II inputs are not assumed interchangeable. No correction factor was fitted.
Daymet precipitation sums the seven days before collection, excluding collection day. It is local gridded precipitation, including snow water equivalent, not a watershed average. All 31 windows are complete. The table screens each reported input separately against the 105-sample Great Lakes reference cohort. Reference ranges are 2.90–92.00% for developed land, 5.65–964.92 people/km² for population density and 0.00–75.77 mm for precipitation (displayed limits rounded). Missing and withheld values are marked unavailable. Approximate inputs are screened as reported. Separate candidate values are excluded. A value within its reference range can still come from a different watershed method or occur in an unfamiliar combination of predictors.
Methods and condition encoding still require review. All Northeast model readiness remains not established. These historical outcomes have already been inspected; they are not an untouched prospective test set. CSV exports include window dates, source links, candidate values and unresolved flags. Microplastic source release.
Case studies on capturing and processing microplastics in water and wastewater.
Original pilot study · Industry context on laundry wastewater and filter residues